Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Spike-RetinexFormer: Rethinking Low-light Image Enhancement with Spiking Neural Networks
Authors: Hongzhi Wang, Xiubo Liang, Jinxing Han, Weidong Geng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Across standard benchmarks, the method matches or surpasses strong ANNs (25.50 dB on LOL-v1; 30.37 dB on SDSD-out) with comparable parameters and lower theoretical energy. Our work pioneers the synergistic integration of SNNs into Transformer architectures for LLIE, establishing a compelling pathway toward powerful, energy-efficient low-level vision on resource-constrained platforms. |
| Researcher Affiliation | Academia | Hongzhi Wang School of Software Technology Zhejiang University Ningbo, China EMAIL Xiubo Liang School of Software Technology Zhejiang University Ningbo, China EMAIL Jinxing Han School of Software Technology Zhejiang University Ningbo, China EMAIL Weidong Geng College of Computer Science and Technology Zhejiang University Hangzhou, China EMAIL |
| Pseudocode | No | The paper describes the architecture and mechanisms (e.g., Figures 1, 2, and 3) but does not include any explicitly labeled pseudocode or algorithm blocks in the main text. |
| Open Source Code | No | The paper does not contain an explicit statement about the release of source code or a link to a code repository. |
| Open Datasets | Yes | We evaluate Spike-Retinex Former on a comprehensive suite of standard benchmarks for LLIE, largely following the protocol of [43]. These include: LOL-v1 [6], LOL-v2 (real and synthetic) [44], SID [45], SMID [46], SDSD (indoor and outdoor) [47], MIT-Adobe Five K [48], and LIME[49] |
| Dataset Splits | Yes | For datasets with paired ground truth, we train a separate model per dataset using the official or widely adopted splits; for RAW datasets, we convert to sRGB via the standard ISP pipeline before computing losses and metrics to ensure comparability. |
| Hardware Specification | Yes | All ablations are conducted on LOL-v1 using a single RTX 3090. |
| Software Dependencies | No | The paper mentions software components like 'AdamW optimizer' and 'LIF neurons' but does not specify version numbers for any libraries or frameworks used (e.g., PyTorch, TensorFlow, Python versions). |
| Experiment Setup | Yes | All experiments share the same backbone and training hyperparameters as in Sec. 3. Unless otherwise stated, we use the AdamW optimizer (base learning rate 2e-4) with cosine annealing (optional warm-up), unrolling the spiking dynamics for T time steps and applying global-norm clipping at 1.0. Each model is trained until the validation performance saturates. We use ζ=6 and (κ, κp)=(1.0, 0.5) in all experiments. The loss combines an L1 image term, a firing-rate regularizer, and a temporal TV on the pre-readout state to suppress flicker without over-smoothing spatial details: L = |Y - J|_1 + λspk Avg Rate(S) + λtv TVt(H(out)[1:T]), with (λspk, λtv) = (10^-4, 5e-4). We therefore adopt T=8 as the default trade-off. |